P(doom) estimates from AI researchers span approximately six orders of magnitude — from Yann LeCun's under 1% to Eliezer Yudkowsky's over 95%. This enormous range reflects genuine, substantive disagreement about the future of AI and the difficulty of the alignment problem. Understanding why experts disagree is essential for anyone trying to form their own view of AI existential risk.
P(doom) Calculator
Estimate your own P(doom) by adjusting probability inputs for key risk factors. See how your assumptions combine.
P(doom) Calculator →Expert Estimates Overview
Here are prominent public estimates from leading AI researchers and industry figures:
| Expert | P(doom) Estimate | Context |
|---|---|---|
| Yann LeCun (Meta) | <1% | Believes current architectures incapable of existential risk |
| Andrew Ng | ~1% | Views AI risk as overblown compared to other risks |
| Yoshua Bengio | ~20% | Based on 50% AGI probability × 40% alignment failure |
| Sam Altman (OpenAI) | 10-20% | "Low but non-zero" in various interviews |
| Geoffrey Hinton | 10-50% | Estimate has increased over time as capabilities advance |
| Paul Christiano (ARC) | 10-20% | Former OpenAI alignment researcher |
| Dario Amodei (Anthropic) | 10-25% | Focuses on catastrophic rather than existential risk |
| Stuart Russell | Significant | Regards risk as serious and motivating, no precise number |
| Nick Bostrom | High | Author of "Superintelligence," no precise public estimate |
| Eliezer Yudkowsky (MIRI) | >95% | Most pessimistic prominent voice, believes alignment nearly impossible |
The AI Impacts Expert Survey
The most systematic data comes from the AI Impacts Expert Survey, which surveys thousands of AI researchers at conferences like NeurIPS. The 2022 survey of 738 ML researchers found:
- Median P(extremely-bad outcome): 5%
- Mean P(extremely-bad outcome): 14.4%
- 10th percentile: ~1%
- 90th percentile: ~50%
This survey is often cited as the "expert consensus" — but the wide spread between median and mean, and between 10th and 90th percentiles, shows that even among experts, disagreement is substantial.
Why Experts Disagree
The enormous spread in estimates is not random — it reflects systematic disagreement about specific questions:
1. AGI Timelines
When will transformative AI arrive? Optimists (often with lower P(doom)) expect longer timelines (50+ years), giving more time to solve alignment. Pessimists expect shorter timelines (5-15 years), leaving insufficient time for safety work. The 2022 AI Impacts survey found median estimates of ~50% chance of AGI by 2059, but individual estimates ranged from never to within a decade.
2. Alignment Difficulty
Is alignment a solvable engineering problem (LeCun's view) or a fundamentally hard theoretical problem (Yudkowsky's view)? This is perhaps the core disagreement. Those who view alignment as tractable tend to have lower P(doom); those who view it as extremely difficult tend to have higher P(doom). Recent progress in RLHF and Constitutional AI has moved some estimates downward, but skeptics argue these methods won't scale to superintelligence.
3. Capability Jump
Will AI capabilities increase gradually (allowing course correction) or suddenly (removing time to respond)? Gradualists argue we'll have warning signs and time to adapt. Sudden-jump theorists worry about "intelligence explosion" scenarios where AI self-improvement rapidly outpaces human response.
4. Societal Response
Will international coordination succeed? Will competitive pressure override safety investment? Optimists about governance tend to have lower P(doom); pessimists about race dynamics tend to have higher P(doom). The current competitive landscape between major AI labs and nations is a significant concern for many researchers.
5. Definition of "Doom"
Experts use different definitions. Some define doom as literal human extinction; others include civilisational collapse, permanent loss of autonomy, or "any outcome significantly worse than a good future." Broader definitions naturally produce higher estimates.
Calculate Your Own P(doom)
Our P(doom) Calculator lets you explore how changing your assumptions about these key variables affects the combined probability estimate. It's a useful tool for stress-testing your intuitions about AI risk and understanding how different experts might arrive at their estimates.
P(doom) Calculator
Adjust probability inputs for alignment failure, misuse risk, timeline, and societal response. Explore how your assumptions combine into a composite estimate.
Open Calculator →